"ITAGAKI Takahiro" <itagaki.takahiro@oss.ntt.co.jp> writes:
> 4. ANALYZE finishes in a short time.
> It is ok that VACUUM takes a long time because it is not a transaction,
> but ANALYZE should not. It requres cleverer statistics algorithm.
> Sampling factor 10 is not enough for pg_stats.n_distinct. We seems to
> estimate n_distinct too low for clustered (ordered) tables.
Unfortunately no constant size sample is going to be enough for reliable
n_distinct estimates. To estimate n_distinct you really have to see a
percentage of the table, and to get good estimates that percentage has to be
fairly large.
There was a paper with a nice algorithm posted a while back which required
only constant memory but it depended on scanning the entire table. I think to
do n_distinct estimates we'll need some statistics which are either gathered
opportunistically whenever a seqscan happens or maintained by an index.
-- Gregory Stark EnterpriseDB http://www.enterprisedb.com Ask me about EnterpriseDB's Slony Replication
support!